Skip to content

DLSlime

Composable communication runtime for AI services

DLSlime is a PeerAgent-centered communication and microservice toolkit for distributed AI systems. It lets applications adopt endpoint transfers, PeerAgent coordination, SlimeRPC, and DLSlimeCache one layer at a time.

Start Here

Installation
Install from PyPI or build optional transports from source. Quickstart
Run endpoint, PeerAgent, cache, and RPC examples. Guide
Use DLSlime components in service-shaped applications. Architecture
Understand how PeerAgent, NanoCtrl, and endpoints fit together.

Core Layers

Layer Use it when
Endpoint API You already control peer placement and metadata exchange.
PeerAgent You want connection setup, memory-region discovery, and stale-state cleanup handled by DLSlime.
DLSlimeCache Multiple PeerAgent clients need a shared RDMA-backed cache service.
SlimeRPC Application logic should call Python services while keeping transport coordination inside DLSlime.

Typical Service Flow

  1. A service starts and registers itself with NanoCtrl as a generic entity.
  2. Each service attaches to a PeerAgent instead of managing transport state directly.
  3. PeerAgents register resource records and memory regions with NanoCtrl.
  4. Clients discover services by kind and scope.
  5. Endpoint objects issue the actual transfer through RDMA, NVLink, Ascend Direct, or the selected backend.